Users’ experience on the internet is positively influenced by recommendation systems, which help them in navigating through a lot of information available to them. The accuracy and relevance of recommendations can be improved only if advanced techniques like GNNs are used due to the increasing complexity associated with user-item interactions. Detailed patterns in preferences and behavior are captured by GNNs which model relationships between users and items. To this end, we present a Position-Aware Neural Graph Collaborative Filtering (PA-NGCF) model with two key differences from previous graph neural network-based recommender systems. In terms of our primary contribution, we have come up with a new way of creating node embeddings that make use of limited rating information from users to adequately capture user-item interactions. Furthermore, during message passing phase in our model, the positional information about nodes within the graph structure is explicitly included enabling richer understanding about the role and significance of each node involved in recommendation process. Two real-world datasets were used for extensive experiments to show how effective our approach was. Our findings indicate that PA-NGCF method effectively reduces the rating prediction error within recommendation systems. This error has been evaluated using the MAE and RMSE metrics. Our findings highlight the potential of the proposed method in increasing the quality of recommendations and addressing cold start problems in sparse datasets, as the reduction in prediction error leads to more accurate ranking of items adapted to user preferences.
